We were relying on FE_DIVBYZERO being turned off when printing
"%f", +-.0
Avoid the whole issue by looking at the sign-bit (in a rough
approximation).
Note that we do not handle gracefully:
printf ("\n%llf\n", -0.0);
printf ("\n%llf\n", 0.0);
nor %Lf for both when NOT cast to long double. Avoiding an FPE due to
broken numbers like these does not make sense to me.
Signed-off-by: Bernhard Reutner-Fischer <rep.dot.nop@gmail.com>
*/
#define isnan(x) ((x) != (x))
-/* Without seminumerical functions to examine the sign bit, this is
- * about the best we can do to test for '-0'.
- */
-#define zeroisnegative(x) ((1./(x)) < 0)
-
/*****************************************************************************/
/* Don't change anything that follows peroid!!! ;-) */
/*****************************************************************************/
if (x == 0) { /* Handle 0 now to avoid false positive. */
#ifdef __UCLIBC_HAVE_SIGNED_ZERO__
- if (zeroisnegative(x)) { /* Handle 'signed' zero. */
+ union {
+ double x;
+ struct {
+ unsigned int l1, l2;
+ } i;
+ } u = {x};
+ if (u.i.l1 ^ u.i.l2) { /* Handle 'signed' zero. */
*sign_str = '-';
}
#endif /* __UCLIBC_HAVE_SIGNED_ZERO__ */